Behavioral System to Detect Injury and Rehabilitation Process in Karate Using Hybrid Model

1.

Introduction

Karate is said to have originated on Okinawa from native Okinawan fighting techniques, as well as Chinese and Japanese martial arts. It was introduced to Japan in the early 20th century, and by the middle of the century, its popularity had grown. As the first global karate governing body, the World Karate Federation (WKF) was established in 1970 and currently has 191 member nations. Karate is getting more and more popular worldwide, and in 2020 it will make its Olympic debut in Tokyo. Now, kata and kumite events with individual and team elements can be seen in karate tournaments. Kumite is free combat with an opponent using the same techniques, whereas kata is the application of prepared techniques without any direct confrontation.

A complicated, hard, and risky sport is karate. These days, as athletes try to be as effective as possible during practice and competition, their injuries often get worse. Numerous indicators point to the fact that sports injuries are on the rise and are a major concern for sportsmen. Technical errors, insufficient intensity, prolonged time, excessive repetitions, or abrupt movements are the main causes of athlete self-injury. Additionally, the likelihood of an athlete suffering from a sports injury increases with the duration of their physical activity and the severity of the ailment. The most frequent wounds suffered in martial arts are cuts, bruises, sprains, and strains. A bone may also fracture. Injury to the knee, ankle, shoulder, and elbow is among the most frequent ones. When practicing martial arts that entail hitting, the hands are especially vulnerable to harm. The lack of ability to “see”, evaluate, and compare harm at the tissue level, according to Flint [1], is the root cause of the variety of injury classifications.

Understanding the traits, causes, treatment options, and prevention of karate injuries is essential to approaching the issue of karate injury prevention, treatment, and rehabilitation. The fundamental properties of karate, as well as its distinct physical features, must also be understood by individuals, as must any potential connections between these elements and injuries. A detailed examination of the physical attributes of karate, the project’s attributes, and the characteristics of injuries is followed by the search for an all-encompassing set of risk assessment procedures.

Studying the factors that assist athletes in reducing IR is crucial. Injuries cannot be completely eliminated in karate since it has not yet been established that impacts that exceed the mechanical resistance of tissues are the primary cause of major injuries. Karate assaults specifically target vulnerable parts of the opponent’s body, so any modifications to the rules of combat set forth by sports federations can only partially lessen the risk of injury. This article, which uses CNN and DNN models from deep learning, suggests a novel behavioral approach to detect injuries and the recovery process in karate in order to allay these worries.

2.

Literature review

The scientific study of statistics and mathematics that allows computers to learn from experience and make better decisions is known as machine learning. A wide range of scientific, medical, and financial domains have benefited from the application of machine learning, including image identification, cancer detection, stock market forecasting, and customer churn prediction [1, 2]. In many domains, including athletics, machine learning is still in its infancy [2, 3]. Sports injuries are attracting the attention of more sports medicine researchers, despite the fact that many of them lack expertise with psychological evaluation methods. For instance, Everhart et al. used 34 psychological component evaluation measures in 152 sports injury treatment outcome studies that included psychological evaluations prior to and following therapy [4].

Machine learning methods were used in 2023 by Dandrieux et al. [5] to estimate the risk of injury in 284 male football players who were members of 16 professional or semi-professional football clubs from three different nations. Three models adaBoost, random forests, and logistic regression were contrasted using a dummy classifier. The findings demonstrated that using these models, it is possible to forecast the likelihood of injury to some degree. Tondut and others [6] in 2023, they make an effort to calculate the risk of injury for 110 athletes who compete in sports that need repeated sprints (bobsleigh, rugby at 7, athletics, and athletics). A decision tree model was trained and refined to predict the probability of an injury based on the variables that were measured. Both the Sports Medicine Diagnostic Coding System and the Sports Injury Disease Classification System were updated in later editions by Orchard et al. to incorporate the new consensus categories [7].

In addition to providing a number code for the effect of injuries on motor function, Bj et al. released the Orchard Sports Injury Classification System, a Spanish translation of the Sports Injury Classification System’s 12th edition [8]. Osteoarthritis in the knee is associated with trauma to the knee. The Inaa et al. study set out to determine the likelihood of KR surgery and to pinpoint medical costs for the 10 to 15 years after a sports injury [9]. Even while previous studies have significantly improved sports injury prevention, there are still glaring inadequacies in the effectiveness of injury prevention. There has been a lot of interest in applying machine learning to sports injuries, and various academics have conducted research in this area. Desmond M emphasized the efforts made by the US Olympic Ski Team to improve training methods and results in competition by utilizing machine learning (ML) and remote sensor telemetry. In the context of motion dynamics, this section examines the potential applications of machine learning [10].

To encourage children’s interest in machine learning, Zimmermann-Niefield et al. offered a number of simple and effective methods for them to use sensors sensibly when exercising [11]. The problem brought up by Pappalardo et al. has attracted a lot of attention from companies and the scientific community since machine learning algorithms can gather massive amounts of data that can be utilized to avoid sports injuries using electronic data. assessing football players’ performance [12].

Numerous machine learning techniques have been used to examine the effects of sports injuries, and Mittal et al. developed a machine learning architecture to monitor these effects in real time [13]. Even though using machine learning to avoid sports injuries improves preventative effectiveness, the field is now too complex for it to be appropriate. Using a hybrid model as an innovation, this research suggests a novel behavioral technique to identify injury and the healing process in karate.

3.

Fundamentals

Two of the most important aspects of the study described in this paper are the concepts of injury detection and the machine learning (ML) techniques utilized to evaluate the dataset; these topics are covered in length in this section.

4.

Our proposed approach

This section introduces our Behavior Karate AI methodology. The scenario of our approach is shown in the Figure 9.

Four primary phases comprise the suggested approach, as seen in Figure 10: data collection, supervised classification, unsupervised feature learning, and the rehabilitation process. First, inconsistencies (Infinity, NaN, and null values) are eliminated. The second uses a deep autoencoder to carry out unsupervised learning. To create a new representation of the characteristics and minimize the dimensionality of the unlabeled data, the auto-encoder is utilized. Based on the results of the first phase, a deep neural network (DNN) or a convolutional neural network (CNN) is then used in the third step to carry out supervised learning classification.

The process of rehabilitation is presented in the fourth step. The ensuing subsections include more information.

Phase 1: Procedure for gathering data. The primary goal of this phase is data preparation and gathering. In our instance, information was gathered through a test administered by the Tunisian Karate Federation. All of the data that we gathered is contained in a file called karate.csv.

Phase 2: unsupervised learning. When numerous features are available, the complexity of the generated model may grow as a result of noise and redundancy, which may impair learner performance. By shrinking the size, these issues can be resolved. The adjustments required to extract the most significant features from unlabeled data inputs are made using a feature extraction technique. Following the production of low-dimensional data, the classifier will be categorized to improve its worldwide efficiency. As an unsupervised neural network approach, we employed an autoencoder at this phase. Its primary goal is to replicate the original input as a reduced, equivalently rebuilt output.

Phase 3: Supervised classification. The newly derived feature space and the class vector from the original dataset will be joined in this stage. Consequently, we are left with a condensed and labeled training set that our preferred classifier can use as input data. Numerous supervised learning methods exist, including decision trees, convolutional neural networks (CNNs), deep neural networks (DNNs), random forests, and neural network classification rules.

We have chosen Deep Neural Network (DNN) and Convolutional Neural Network (CNN) as our two supervised classifiers for this challenge. In this instance, we decided to use the online Google Colab workspace. We used two models to assess the CNN and DNN datasets and the autoencoder for the feature reduction method to evaluate our approach.

Phase 4: The rehabilitation process. In this stage, if there are injuries, the doctor proposes a strategy for rehabilitation. Injury severity was assessed based on days without physical activity. We have established four levels of severity:

• Level 1: No suspension of activity.

• Level 2: suspension of less than 8 days;

• Level 3: 8 to 30 days of unavailability.

• Level 4: 30 days or more of interruption or requiring hospitalization, cast or surgical care.

5.

Material and method

6.

Evaluation and discussion

Accuracy, Precision, Recall, F1-score, and Area under the receiver operating characteristics (ROC) curve are used to compare and assess the suggested approach. We employed the macro and micro averages of Recall, Precision, and F1-Score in this work. Score for multi-class classification. All of the metrics listed above can be produced with the use of the confusion matrix (CM). Table 2 demonstrates CM is designed for binary classes, but it can be extended to many classes.

In Table 2, True positive (TP) refers to the amount of class_pos data predicted that actually belongs to class_pos, True negative (TN) refers to the amount of classneg data predicted that actually belongs to class_neg, False positive (FP) refers to data predicted as class_neg but actually belongs to class_pos, and False negative (TN) refers to data predicted as class_neg but actually belongs to class_pos. The evaluation metrics are calculated using the terms listed above.

Accuracy: It is the proportion of cases that are correctly classified to all occurrences. It is also known as detection accuracy, and it is only a valuable performance metric in balanced datasets.

Precision (P): To all the samples predicted as Attacks, it is the ratio of accurately predicted Attacks to all the samples.

Recall (R): It represents the proportion of samples that are indeed attacks to samples that were accurately identified as attacks. It also goes by the name “Detection Rate.”

F-measure (F): Its definition is the Precision and Recall harmonic mean. Stated differently, it is a statistical method for assessing a system’s correctness by taking into account the system’s recall and precision.

To assess the effectiveness of our suggested strategy, we prepared a database from the quiz carried out within the Tunisian Karate Federation. We created a csv file “karate.csv”, which contains all the information about the athletes: name, age, date of birth, number of injuries,… This file was an input into our model. In our instance, we decided to work on a Google Research product called Google Colab. Colab respects data privacy rules by enabling anyone to write and execute any Python code through a browser. It is a very good setting for teaching, learning from data, and machine learning. Technically speaking, Colab is a hosted Jupyter notebook service that offers free access to computer resources, including GPUs, and doesn’t require any configuration. For supervised learning, we used CNN and DNN models, and for unsupervised learning, we used Autoencoder. We’ve run multiple tests on our model. The optimal experimental outcomes of our CNN and DNN models are displayed in Table 3.

An attempt should be made to better understand the relationship between training and competition load and injuries in order to quantify tiredness and recovery, better comprehend adaptability to training programs, and lower the risk of illness and injury. Several data points have been combined in karate for analysis and damage prediction. But it wasn’t until lately that the right statistical methods were applied to examine the accessible data set.

The intricacies of the relationship between player load and injury are now more understood because to machine learning’s advancements in autonomous and interactive data analysis. Our methodology has already been put to the test in the intrusion detection domain, and we have presented the findings at two international conferences.:

• “Behavioral System for the Detection of Modern and Distributed Intrusions based on Artificial Intelligence techniques: Behavior IDS-AI” in the 6th International Joint Conference on Advances in Computational Intelligence IJCACI 2022 [16].

• “Behavior Intrusion Detection System Using SVM And CNN” in the Congress on Smart Computing Technologies (CSCT 2022) [17]. In this work, we chose to apply our model in the field of karate to detect injuries, and we noticed that our model is efficient and gave us good results.

We compare our approach here with several works that are worked on maching learning: Vallance et al. [18], Naglah et al. [19], Kamakis et al. [20], Jauhiainen et al. [21], Henriquez et al. [22], Lovdal et al. [23], Van Eetvelde et al. [24], Taborri et al. [25], Majumdar et al. [26]. Table 4 compares our best findings with other writers’ findings regarding sports injuries.

Figure 13 presents an accuracy comparison between the proposed Behavior Karate-AI model and other benchmark models, demonstrating that Behavior Karate-AI has a greater accuracy than the other models. The suggested Behavior karate-AI model beats its comparison peers in terms of accuracy, reaching 99.67% and 99.66% for DNN and CNN, respectively. Table 3 compares the proposed model and existing literature models in terms of accuracy. In certain instances, the findings are aggregated and do not reveal information about the injuries found, and some of the models under comparison do not employ crossvalidation. The overall result demonstrate the value of using Behavior karate-AI to detect injuries in the data set, even though more research is need to compare our findings with those of the literature. By using these techniques, one can maintain the results with those of other comparable works while somewhat reducing the performance gained.

7.

Conclusions

By using predictive models such as CNN and DNN, existing incorrect predictions due to excessive complexity are challenged and interactions between variables are minimized. In karate, such algorithms can be used to detect damage. This helps to identify factors of sport-related karate injuries in the general population. Increased awareness, the existence of legislation mandating the use of protective gear in high-risk sports, and encouraging players to wear protective gear regularly can all help reduce the number of karate injuries. Athletes should be educated about the dangers of karate injuries and their consequences.

This article proposes a new behavioral system for detecting injuries and the rehabilitation process in karate using a hybrid model. We evaluated our strategy using two models to analyze the CNN and DNN datasets, as well as the autoencoder for the feature reduction method. Empirical research demonstrates that the suggested model can produce reliable outcomes. Indeed, our model’s accuracy for DNN is 99.67%, while for CNN it is 99.66%. Future research will include more potent tactics to raise the detection rate and precision of our approach.

Acknowledgments

We would like to express our sincere gratitude to the Tunisian Karate Federation (FTKDA, ftkarate.com) that have contributed to the publication of this research paper.

Conflict of interest

The authors have no conflicts of interest to declare that are relevant to the content of this article.

Competing interests

The authors declare that they have no competing interests.

Financial and non-financial interests, funding

No funding was received.

Author contributions

The authors contributed equally to this work.

Research data policy and data availability statements

Data sharing apply to this article because a dataset was generated and analyzed during the current study.